An SSD’s DRAM cache never held your game files. It held the map the controller uses to find them — the logical-to-physical (L2P) translation table that turns the block address the operating system asks for into the physical location on NAND where the data actually sits. That map has to be consulted on almost every I/O, which is why “always buy a drive with DRAM” was useful advice for a decade. The advice is now contested, and 2026 is a reasonable year to settle it: the drives changed, the API that streams game assets changed, and the review record is finally deep enough to separate one workload from another.
What changed: Host Memory Buffer (HMB) lets a DRAM-less NVMe drive borrow a slice of system RAM — commonly 64 MB, sometimes as little as 40 MB — to cache its mapping table, closing most of the gap a dedicated DRAM chip used to buy. Why it matters: For loading a pre-warmed level, a modern DRAM-less drive is hard to tell apart from a DRAM-cached one. For installing, patching and sustained writes, it is not. Who should care: Anyone putting a second, high-capacity games drive in a budget or mid-range build, and anyone who patches large titles often.
What the DRAM chip is actually doing
Every SSD controller maintains that L2P table. On a drive with DRAM, the table is cached in fast onboard memory; on a DRAM-less drive the table lives in NAND. Controller vendor Innodisk’s explainer spells out the cost: a read may need “two separate NAND accesses: one to fetch the mapping, and another to retrieve the actual data,” and random reads suffer most because those lookups are scattered rather than sequential.
HMB is the correction. Added in the NVMe 1.2 specification, it “allocates a portion of the host DRAM for the SSD,” as Kyusik Kim and Taeseok Kim put it in their 2020 PLOS ONE study, which measured the feature on commercial drives. Innodisk describes the handshake: the SSD tells the host it supports HMB and requests a size, the host checks available memory, and if it agrees it grants access over the PCIe bus. The PLOS study’s conclusion matters more than its mechanics: the DRAM-less drives it tested “mainly exploit the host memory buffer as an address mapping table cache rather than a read cache or write buffer.” HMB buys back the map — not a write cache, and not a read cache. The same study found the DRAM-less drives “clearly exhibit worse I/O performance than SSDs with internal DRAM” across many patterns, but that enabling HMB recovered a large part of the loss, particularly for random reads.
The size of the loan is small and fixed by the controller. TechPowerUp records Phison’s E31 using 64 MB for its mapping tables, the same figure its database lists for the Silicon Power UD90’s HMB, and it found the MaxioTech MAP1602 capped at 40 MB even on a 4 TB drive. That capped HMB is likely why MAP1602 drives post lower 4K random writes than their sequential scores suggest — a link TechPowerUp itself draws.
Two caveats keep the picture honest. First, ServeTheHome’s explainer is blunt: “While many in the industry like to say that HMB SSDs are as fast as their DRAM-equipped alternatives, they are not.” HMB is a value feature that removes a DRAM package and shrinks the PCB. Second, “DRAM-less” is not one thing. TechPowerUp’s Acer Predator GM7 4 TB review states plainly that the category “has a decent bit of variation within it. Performance very much depends on the controller and how it handles that configuration.” Controller firmware, not the DRAM bit, is the variable that moves.
The comparison, workload by workload
This is the table the rest of the article unpacks. “Equal” means the difference is inside the run-to-run noise of the benchmark that measured it; the evidence column names that benchmark.
| Workload | DRAM-less (HMB) | DRAM cache | Evidence |
|---|---|---|---|
| Cold level load, pre-warmed library | Equal | Equal | TechSpot’s 15 game loads showed no meaningful difference between PCIe NVMe drives |
| Reloading a save in play | Equal | Equal | TechSpot: differences “diminish” on reload; the game is bottlenecked elsewhere |
| In-game DirectStorage asset streaming | Near-equal | Near-equal | TechSpot: DirectStorage titles gave PCIe 5.0 only rare single-digit edges |
| Verifying / patching a game (many small files) | Slightly slower | Faster | TechSpot: verifying F1 25 (83 GB, 36,000 files) — DRAM drives 8–11% faster |
| Installing or copying a large game | Slower | Faster | TechSpot: 123 GB copy to the A440 finished 38% sooner than to the A440 Lite |
| Writing past the SLC cache | Much slower | Slower | TechPowerUp fill tests; How-To Geek’s QLC example drops to ~400 MB/s |
| Shader-compilation-heavy first load | Untested directly | Untested directly | No game benchmark isolates it; the workload is unmeasured directly |
| Filling the drive toward capacity | Larger penalty | Smaller penalty | LTT Labs: ~10% shorter FFXIV load times below 50% fill on 1 TB drives |
Where DRAM-less still loses
Sustained writes past the SLC cache. This is the oldest and clearest gap, and it is a firmware-and-cache story more than a DRAM story. TechPowerUp’s Kioxia Exceria Plus G4 — a DRAM-less Phison E31 drive — writes at 6 GB/s until 408 GB have landed, then falls to 1.4 GB/s, then to about 700 MB/s, filling the whole 2 TB at 1060 MB/s on average. The DRAM-less Acer Predator GM7 4 TB is a better case: a 615 GB SLC cache, around 2000 MB/s once exhausted, and a whole-drive fill at 1533 MB/s, which TechPowerUp called the fastest result in its test group for any DRAM-less drive. How-To Geek’s example sits at the other end: a DRAM-less QLC drive (the WD Blue SN5100) drops to roughly 400 MB/s once its SLC cache fills. DRAM-less plus QLC is the combination that hurts.
Installing and patching. These are write-heavy and, in the patch case, many-small-file-heavy. TechSpot’s 2025 suite found that copying a 123 GB game to the DRAM-cached A440 took 38% less time than to the DRAM-less A440 Lite — a gap that mostly reflects SLC cache size (around 200 GB versus 48 GB), with “over half the game” written directly to TLC on the Lite. The single result where a DRAM cache made a measurable difference anywhere in that suite was verifying F1 25, an 83 GB game spread across 36,000 files: the three DRAM-equipped PCIe drives finished 8–11% faster than the three without. Small files are where the mapping table gets hit hardest.
DirectStorage asset streaming. Microsoft’s own documentation sets the design target at “50,000 requests per second” using at most 10% of a single CPU core, and describes desktop DirectStorage as delivering “multiple gigabytes a second of small (for example, 64kb) data reads with minimal CPU overhead.” That is a small-request regime, which sounds like the mapping-table workload — but TechSpot’s DirectStorage titles (Assassin’s Creed Shadows, The Last of Us Part II, Spider-Man 2) showed no meaningful gap between PCIe NVMe drives, and only rare single-digit wins for PCIe 5.0. The API’s own guidance discourages the serial dependency chain that would expose drive latency: “Don’t use DirectStorage with dependency change assets,” the developer guidance warns, because waiting 8 ms for a single 512-KiB read works out to 64 MB/s. DirectStorage is built to keep many requests in flight so the drive’s queue, not its DRAM, stays busy. For the PCIe-generation side of that argument, see our DirectStorage and PCIe 5.0 benchmark breakdown.
Shader-compilation-heavy first loads. This is the workload with the weakest direct evidence, so treat it as a pattern rather than a measurement. Microsoft’s GDC 2026 session on Advanced Shader Delivery calls long shader compilation and D3D12 shader stutter “two of the biggest problems in PC gaming,” caused by compiling shaders at runtime, and describes precompiled shader databases delivered for a player’s specific hardware. Compilation itself is CPU and driver work, not storage work. What touches storage is the shader cache it writes and re-reads — and Microsoft notes that some titles carry “such a large amount of pipeline state objects (PSOs) to the point that most engines cannot enumerate them.” The closest measured proxy in the record is TechSpot’s 36,000-file verification test — the one place a DRAM cache showed a repeatable edge. Nobody has published a DRAM-less-versus-DRAM game benchmark that isolates shader compilation, so the honest statement is that the access pattern looks unfavourable and the direct test has not been run.
Where DRAM-less does not lose
Loading a level you have already cached is the case that made HMB credible. Across TechSpot’s 15 titles — including DirectStorage ones and cold loads into large open worlds — PCIe NVMe drives of every specification loaded close together, and the drive’s specification was not what separated the quick from the slow. On reloads from inside a game the spread narrows further, because the drive is no longer the bottleneck: the CPU has to process the level and the GPUGPU. Graphics Processing Unit — the chip that renders the game's visuals; the main driver of framerate and image quality. has to upload the assets. This is the “sequential load of a pre-warmed asset” case, and here the DRAM binary really does stop mattering. LTT Labs’ fill-level study shows the same shape from the other direction: pushing a drive from half full to completely full changed game frame rates by nothing measurable, and moved FFXIV scene load times by “less than a 10% degradation.”
Which numbers here are not game numbers
Benchmark figures and game figures answer different questions, and this topic is full of the mismatch. Be specific about which is which.
- “Up to 14,500 MB/s” is a peak sequential read at deep queue depth on a fresh drive. No game reads that way; TechSpot found PCIe 3.0 and PCIe 5.0 near-identical in its DirectStorage titles.
- “1.2–2 GB/s sustained write” comes from TechPowerUp’s fill test — a single-threaded stream of 1 MB blocks written to an erased drive until the SLC cache empties. It is a torture test for the cache, not a game load, and its numbers rise sharply once the stream pauses and the cache refills.
- SLC cache sizes (64 GB on the UD90, 408 GB on the Exceria Plus G4, 615 GB on the GM7) are dynamic: they shrink as the drive fills, because the drive reclaims pseudo-SLC cells as TLC or QLC. A cache measured on an empty review sample is the best case.
- TweakTown’s “3DMark SSD Gaming Test” numbers are a trace — a recording of game file access replayed for consistency, which TweakTown itself notes is “the same as running the actual game, just without the inconsistencies.” It is closer to a game than CrystalDiskMark, and still not a game.
- The PLOS ONE figures are fio laboratory measurements on PCIe 3.0 drives, published in 2020. They establish the mechanism, not current performance.
The real game numbers in this article are TechSpot’s level-load and Steam verify/copy timings and LTT Labs’ FFXIV benchmark. Everything else is a proxy.
What it means when you buy
For a games drive that mostly loads and rarely writes, a DRAM-less HMB drive is a defensible choice, and the market now pushes you there anyway: Phison’s own COMPUTEX 2025 release positions the DRAM-less E31T for “ultra thin laptops and handheld gaming devices” — the small form factors where PCB room is tight. Tom’s Hardware’s July 2026 buying piece reaches the same practical conclusion from the other side: patching, downloading and copying use the drive’s write speed, but the effect on actual gameplay is “typically very minimal,” so “you can use anything from PCIe 3.0 to PCIe 5.0.”
What should actually drive the decision is not the DRAM bit but three other things: the controller and its firmware (TechPowerUp’s point about variation inside the DRAM-less category), the SLC cache size relative to the games you install — a 48 GB cache against a 123 GB game is a slow copy — and the flash type. DRAM-less TLC with a large cache behaves very differently from DRAM-less QLC with a small one. If the drive will also hold your operating system, or if you move 100 GB-plus games between drives regularly, the DRAM-cached drive still earns its premium. If it is a second library drive that mostly reads, it does not.
Next read
The storage side of a build only pays off if the rest of the machine can feed it. Our 16GB VRAM baseline report covers the memory that actually constrains modern asset streaming, and the 3D V-Cache CPU comparison covers the other half of the load-time budget. For how the direct-storage path is landing in shipping games, see DirectStorage in 2026: engines, games and load tests, and for the generational drive comparison that started this question, PCIe 5.0 vs Gen4 NVMe load times.